Are Boardrooms Building AI Sandcastles?

Executives across the world are pouring millions into flashy AI pilots, chasing investor applause and headlines. Yet most AI projects have failed to move the profit needle. The illusion of ‘AI-first’ strategies gives rise to a sobering reality – back-office automation may quietly be the true engine of value.

Maya, CFO of a medium-sized logistics company winced at the numbers in front of her, like a general peering out into a storm. Their CEO has demanded an ‘AI-first’ quarter: a novel, fancy customer-facing chatbot, an award-winning white-paper-generator and an ensuing 10% reduction in the workforce within a year. The board cheered, investors smiled and the company contracted a cloud vendor on a six-figure contract.

Six months later, Maya had heatmaps, deployment logs and an uneasy heart: customer satisfaction was no longer increasing the way it did at the start, costs had increased and any time savings were confetti.

That should read familiar.

The AI Hype Cycle Hits the Boardroom

We are currently in the “manic installation” phase of a new and evolving paradigm, according to many experts. Huge amounts of money are being thrown into the visible, headline grabbing aspects of AI. Markets have offered – and are still offering – eye-watering valuations to AI-adjacent companies with price-to-earnings ratios of some U.S. tech stocks nearly 40 times – some even higher.

Yet, are the basics of value creation being overlooked? A sobering body of research suggests – yes, and that the tide may, in fact, soon turn. Most AI pilot programs are not getting the profit needle moving. A widely-publicised survey from MIT’s NANDA initiative released mid-August found 95% of GenAI pilots were bringing no material gain to the company. But – is this a judgement on the science itself, or a judgement against misdirected ambition?

Today’s rising swell of discontent around AI underperformance is likely to do with expectations. When introduced across the world by major tech companies, AI was touted to be the biggest game-changer of this generation, many even likening it to the industrial revolution. Executive teams were attracted to the illusion of general intelligence – a panacea that will reason, decide and be insightful overnight.

Instead, what they’ve received – particularly in the form of Large Language Models (LLMs) – are powerful pattern-matching predictive engines that must be carefully scaffolded: clean data, domain tuning, governance and ongoing engineering. That AI will convert hype to revenue magically, at this point, looks rather unlikely.

Currently, AI users are increasingly finding it is good data, clear objectives and operational discipline that can turn data into incremental advantage, not much more.

Second, organizations more often than not go after the shiny and firmly hold a fear of missing out to retain a competitive advantage. A public chatbot redraws investor stories and develops a press cycle, with back-office automation considered more practical, boring and behind the scenes. But it is the latter where real return on investment is currently being accrued, and often silently.

These small-scale templates, scripts and systems can be summarized as a claim (drafting, scraping, normalizing reports or contracts etc), but also scaled down to individual employee efficiencies (using an LLM to draft responses to a customer service bot, a macro that auto-fills a report template, a junior lawyer using a customized contract extraction customization). In contrast to flashy pilots, they are specific to workflows, and are gauged by time-saving or error-reducing, as opposed to headline impressions.

Third, “intelligence” is costly. Infrastructure, model licensing, data wrangling, regulatory compliance and the alliance of talent required to ensure these systems are kept on the straight and narrow are not mere budgetary add-ons, they require a major cost restructuring. Lots of projects fail because executives do not invest in the plumbing: data pipelines, observability, monitoring and human-in-the-loop review. When you look at AI and think of it, say, as an ongoing operating model not a SaaS subscription you purchase, it will, at least right now, cost more in subscriptions and failures than it will bring benefit.

From Sandcastles to Solid Ground

So what is a profitable, realistic AI programme?

Companies must be ruthless, or at least liberating, with problems – and not models. AI needs to be employed in a way that is tightly constrained and for repeatable tasks. All components of inputs and desired outputs need to be clearly understood. Anomaly detection in well-instrumented systems, document classification, invoice reconciliation, and so forth: these are what will pay off in the near-term.

Second, stress must be placed on integration with existing systems instead of pure novelty. A feature that saves two hours in reconciliation and feeds that data into a live dashboard is impact that will be measured and appreciated long after the glitz of a chatbot on the home page has faded.

Third, executives need to make it a point to invest in governance and operationalization right at the offset. Monitoring, reviewing performance indicators and human supervision must be activities for the first sprint and not the ninth.

—

Maya’s company eventually pivoted. Their public chatbot project was put on the backburner – they worked to automate freight invoice adjudication and feed that process into the commission calculations of the sales team instead. The result? Sparkier cash application, fewer disputes and an observed jump in their operating margin. It was tedious, microanalytical and nothing like the glamour of press releases. The bigger market lesson is equally evident: valuations based on future potential alone without an explanation of how the same will be monetized in the near-term is essentially building sandcastles in the air. The process of AI transformation across verticals is a process that will take place over the decades, but the narrative of those who make money in the short-term will be one of steady, unexciting execution. The AI champions of tomorrow will focus capital on specific data cases today.